Evaluation of High-Density EMG Feature Extraction and Selection to Recognize Lower Limbs Movements for a Rehabilitation Exoskeleton

The purpose of this study is the evaluation of HD-EMG feature extraction andselection for the classification of Lower Limbs Daily Living Movements for the control of arehabilitation exoskeleton. The movements considered in the experimental protocol were:Stand to sit, sit to stand, stair ascending, stair descending, a gait cycle, rest in uprightposition and rest in a sitting position. In particular 17 features, divided in time domainand frequency domain, have been taken into account. Monopolar signals were recordedfrom 7 muscles using 4 matrices of 32 electrodes placed on Rectus Femoris, Vastus Medialis,Vastus Lateralis and Tibialis Anterior and 2 matrices of 64 placed on GastrocnemiusMedialis and Semitendinosus & Biceps Femoris. Single differential EMG signals were used,and a bad channel selection was performed to remove those channels without informations.Nine healthy and voluntary subjects were involved in the experiments. For the selectionof the best features subset the computational time, the wrapper approach based on aLDA classifier and the filter approach based on informations about classes separability (Jindex) were considered. In particular, both features in frequency domain, Mean Frequency(MNF) and Median Frequency (MDF), were excluded immediatly because required toomuch time (1,5s) to be processed. Following a decision rule, a region of interest (ROI)is defined, person per person, as the region with a classifier accuracy greater than 92%and a J index value greater than 70% of the feature with the maximum J index. OnlyWillison Amplitude (WAMP) and Number of Turns (NT) were inside the ROI for all thesubjects and therefore considered as the best subset.